DSPM: Dynamic Semantic Patch Memory for Token-Efficient Long-Dialogue LLM Context Management

Large language models (LLMs) deployed in extended, multi-turn dialogue settings face a fundamental scalability bottleneck: raw conversation histories grow without bound, rapidly exhausting fixed context windows and inflating inference costs. Existing mitigation strategies -- sliding-window truncation and monolithic LLM summarization—achieve token reduction at the expense of critical semantic fidelity. We present Dynamic Semantic Patch Memory (DSPM), a structured, seven-technique compression framework that decomposes conversational memory into typed semantic patches and maintains a token-budget-constrained context through a pipeline of deterministic and utility-driven operators. DSPM achieves a mean Token Reduction Rate (TRR) of 82.4% ± 4.21% across seven heterogeneous technical dialogue scenarios, surpassing the 55% and 60% design targets, while retaining a mean consistency score of 3.57/5.0 relative to full-history baselines. Critical constraints and decisions are preserved through a guaranteed retention mechanism, yielding a mean Critical Retention Rate (CRR) of 94.2%. All experiments are reproducible on commodity hardware using free-tier API access, demonstrating the accessibility of the approach.